This analysis was set in motion after running across extensive wildfire data on Kaggle, published by the national Fire Program Analysis. These data are the accumulation of federal, state, and local reporting systems for wildfires and represent 1.88 million wildfires from 1992-2015. Each fire is geo-located by latitude and longitude, which makes mapping them possible and essential. Here, I replicate and expand upon an article by the Civis Journal that uses R’s leaflet package to map publicly available federal grant data. In an effort to explore the geographic distribution of the most destructive wildfires, I restrict the data to wildfires that burned in excess of 100 acres in California during 2015.

One important caveat is that wildfires are not unequivocably detrimental. In the right environment, wildfires can promote healthy ecosystems, clearing out disease and allowing an area to regain biodiversity. However, climate scientists are increasingly concerned with the effect of climate change on wildfire propagation and intensity. Understanding the primary drivers underlying wildfire behavior will be crucial to fire prevention and containment.

The vast majority of the 1.88 million wildfires are less than 100 acres so our final dataset is comprised of 151 fires. This number is small enough to create a comprehensible map showing points for each fire.

Since preliminary data exploration is mandatory for any data project, I’ll first walkthrough some summary statistics of relevant variables within my restricted fires dataset. I do display most of the code in this analysis but it can be found in full on github.

How big are these fires?

#order of the fire sizes
fire_order <- c("100-299 acres", "300-999 acres", "1000-4999 acres","5000+ acres")

#create ordered, articulate fire size variable
fires <- fires %>%
  mutate(
    fire_size_category = case_when(
      FIRE_SIZE_CLASS=="D" ~ "100-299 acres",
      FIRE_SIZE_CLASS=="E" ~ "300-999 acres",
      FIRE_SIZE_CLASS=="F" ~ "1000-4999 acres",
      FIRE_SIZE_CLASS=="G" ~ "5000+ acres"
    ),
    fire_size_category=factor(fire_size_category, levels = fire_order)
  )
#make table
fires %>%
  count(fire_size_category)
## # A tibble: 4 x 2
##   fire_size_category     n
##   <fct>              <int>
## 1 100-299 acres         68
## 2 300-999 acres         30
## 3 1000-4999 acres       26
## 4 5000+ acres           27

Almost half of the 151 burned 100-299 acres, however there were 53 fires in 2015 that each burned over 1000 acres of California. Potentially more concerning were the 27 fires that grew to a size 5000 or more acres.

#histogram of distribution
fires %>%
  ggplot(aes(FIRE_SIZE)) +
  geom_histogram() +
  theme_bw() +
  xlab("Size of fire (acres)")

Histogram of the fire size (in acres) reveals a very sharp left skew with a long right tail. The 1 or 2 fires of over 150,000 acres are reducing the resolution that we can see the less extreme parts of the data. I log-transform the fire size variable in order to reduce the influence of extreme values in the display of the distribution.

#log transform to reduce influence of upper outliers
fires %>%
  ggplot(aes(log(FIRE_SIZE))) +
  geom_histogram() +
  theme_bw() +
  xlab("Log-transformed size of fire (acres)")

Even when log-transformed, fire size distribution has a heavy right tail. The main takeaway here is that the burn area of most fires was less than 10,000 acres, with a few fires growing to in excess of 50,000 acres.

When do most fires start?

#aggregate and plot the time trend of number of fires
fires %>%
  mutate(startDate=as.Date(paste0(FIRE_YEAR, "-", DISCOVERY_DOY), format="%Y-%j"), #make date variable
         month=factor(months.Date(startDate), levels = month.name)) %>% #make month variable
  group_by(month) %>%
  count() %>%
  ggplot(aes(month, n)) +
    geom_point() +
    xlab("Fire Start Month") +
    ylab("Total fires w/ 100+ acres burned") +
    theme_bw()

Unsurprisingly, the vast majority of fires started in the summer months of June, July, August. Notably, though, the shoulder season months of April, September, and October contained 30 fires combined, indicating that fire survelliance cannot end once summer does.

When do the most destructive fires start?

#summarize acres burned
fires %>%
  mutate(startDate=as.Date(paste0(FIRE_YEAR, "-", DISCOVERY_DOY), format="%Y-%j"), #make date variable
         Month=factor(months.Date(startDate), levels = month.name)) %>% #make month variable
  group_by(Month) %>%
  summarise(fire_size=sum(FIRE_SIZE)) %>%
  ggplot(aes(Month, fire_size)) +
    geom_point() +
    xlab("Fire Start Month") +
    ylab("Number of acres burned") +
    theme_bw()

Also fairly predictably, the monthly sum of number of acres burned closely tracks the monthly fires observed above. Fires that started in July caused the greatest destruction by far, totaling to over 500,000 acres burned.

What are the causes of the fires?

#table of counts
fires %>%
  count(STAT_CAUSE_DESCR, sort = T)
## # A tibble: 9 x 2
##   STAT_CAUSE_DESCR      n
##   <fct>             <int>
## 1 Lightning            54
## 2 Miscellaneous        34
## 3 Missing/Undefined    33
## 4 Equipment Use        10
## 5 Debris Burning        9
## 6 Arson                 7
## 7 Smoking               2
## 8 Powerline             1
## 9 Structure             1
#table of acres burned by each cause
fires %>%
  group_by(STAT_CAUSE_DESCR) %>%
  summarise(total=n(), acres_burned=sum(FIRE_SIZE)) %>%
  arrange(-acres_burned)
## # A tibble: 9 x 3
##   STAT_CAUSE_DESCR  total acres_burned
##   <fct>             <int>        <dbl>
## 1 Lightning            54      456088.
## 2 Missing/Undefined    33      177473 
## 3 Miscellaneous        34       78218 
## 4 Structure             1       69363 
## 5 Equipment Use        10       46404 
## 6 Arson                 7        6862 
## 7 Debris Burning        9        1624 
## 8 Powerline             1         860 
## 9 Smoking               2         298

The table above shows the counts of the causes of wildfire as well as the number of acres burned for each cause. Wildfires caused by lightning were the most prevalent and led to the largest number of acres burned of any cause. Powerline, arson, and smoking caused wildfires, while they garner great media attention, were actually some of the least prevalent causes, only spawning 10 fires combined in 2015. The cause of the fire does appear to be related to total acres burned, which would be useful information if we wanted to build a model to predict fire size. Maybe that’ll turn into good fodder for a future post.

Mapping the fires

Now I finally turn toward mapping the fires. I downloaded a shapefile of california county boundaries from State of California Department of Technology website.

The map above shows the spatial distribution of the 2015 California fires after overlaying the points on top of the shapefile. The points are colored by the cause of wildfire, with lightning cause wildfires clearly clustering in Northern California.

#create label for points
fires <- fires %>%
  mutate(startDate=as.Date(paste0(FIRE_YEAR, "-", DISCOVERY_DOY), format="%Y-%j"), #make date variable
         start_month=factor(months.Date(startDate), levels = month.name),
         endDate=as.Date(paste0(FIRE_YEAR, "-", CONT_DOY), format="%Y-%j"),
         end_month=factor(months.Date(endDate), levels = month.name)
  ) %>%
  mutate(fire_label = paste0(
    "Cause: <b>", STAT_CAUSE_DESCR, "</b><br/>",
    "Burned <b>", FIRE_SIZE, " acres </b><br/>",
    "Start: ", start_month, "<br/>",
    "End: ", end_month)
  )

#color pallete
pal <- colorFactor(
  palette = viridis_pal(begin = .95, end = .4, option = 'A')(3),
  domain = fires$fire_size_category
)

#add tiles
leaflet() %>% 
  addTiles() %>% 
  addPolygons(data = ca_shape,
              color = 'white',
              weight = 1.5,
              opacity = 1,
              fillColor = 'black',
              fillOpacity = .8,
              highlightOptions = highlightOptions(color = "#FFF1BE", 
                                                  weight = 5),
              popup = ~NAME) %>% 
  addCircleMarkers(data = fires,
                   popup = ~fire_label, #create a label with fire cause, start month, acres burned, time burned
                   stroke = F,
                   radius = 6,
                   fillColor = ~pal(fire_size_category),
                   fillOpacity = 2) %>%
    addLegend(data = fires,
            pal = pal,
            values = ~fire_size_category,
            title = "Fire Size")

This map shows the substantial power of the leaflet package maps over static maps. I can explore not only spatial distribution of fires but also find specific forests or towns that each wildfire touched, since the points and shapefile are overlayed on OpenStreetMap. I added labels to each point to provide more information regarding each wildfire.